AI Bias & Fairness Testing demo
Run the unfair-discrimination tests a regulator expects on any pricing, underwriting or claims model, and get the written explanation with them. 1:14
What the demo shows
A bias testing solution that runs unfair discrimination tests on pricing and claims models.
Six agents read the model, estimate group membership for the test portfolio, run the six tests against the carrier's thresholds, find the inputs that stand in for a protected class, refit less discriminatory alternatives and draft the test record with every number cited; the model risk validator signs.
Transcript
- AI Bias & Fairness Testing checks ensures pricing and claims models for unfair discrimination.
- New York expects these tests after every model change, with a written explanation.
- Kestrel Point's new auto rating model goes on the bench, six tests.
- For fail, black drivers reached the preferred tier at 30%, white drivers 42%.
- The proxy detector shows that zip code, stands in for race.
- Switch zip code off and the gap falls under the threshold.
- But accuracy slips so the agents refit fair alternatives.
- A density banded territory passes all six tests and keeps 99% of the lift.
- The record cites every number and a model risk validator signs it.
- Built on Botminds: agents, data, screens, and governance on one platform.
Transcribed from the film’s narration. Captions are on the player.